MRHI
View demo → Request access
Capabilities · v0.1 MVP

Everything MRHI ships with.

The complete capability surface — from real-time vector search and metadata filtering to point-in-time recovery, sharding, and cluster operations. The internals stay ours; the speed becomes yours.

Request access Try the public demo

Metadata filtering

  • Equality shorthand plus $eq · $ne · $in · $nin
  • Range and presence: $gt · $gte · $lt · $lte · $exists
  • Array / string predicates: $contains
  • Boolean composition: $and · $or · $not
  • Fast indexed-filter path for common predicates
  • Correctness-preserving fallback for any predicate

Vector formats & metrics

  • Dense formats: float32 · float16 · int8 · int4
  • Packed binary embeddings with exact Hamming distance
  • Distance metrics: cosine · L2 · Hamming
  • Tested dimensions: 64 → 4,096
  • Benchmarked to 100,000,000 vectors

Three engines, one index

  • CPU — x86 · ARM
  • CUDA — NVIDIA H100 / A100 / L40
  • Metal — Apple M-series
  • Same API, same recall envelope on every engine
  • 99.94–100% recall held across all published runs

Persistence & durability

  • Snapshot persistence
  • Write-ahead durability for per-write crash recovery
  • Pure in-memory mode for benchmark and ephemeral workloads
  • Same-process cold-storage switching via unload / load

Backup, restore & point-in-time recovery

  • backup / restore / manifest CLI
  • Point-in-time recovery: --until-time · --until-seq · --until-marker
  • WAL archiving with archive-wal
  • Named recovery points with recovery-mark
  • Retention with prune-backups --keep-last · --keep-days

Scale architecture

  • Segmented indexes that grow without rebuild downtime
  • Local sharding with deterministic ID routing
  • Segmentation and sharding compose for large indexes
  • Per-shard live-id tracking keeps deletes and upserts crisp

Distributed deployment

  • Leader-follower shard groups, or primary-only nodes
  • Durable replication with follower acknowledgements
  • writeAck: primary | quorum | all
  • Binary HTTP payloads for vector-heavy write and search routes
  • Stale-write fencing rejects writes routed to old primaries

Cluster operations

  • Quorum-acknowledged cluster state with leadership leases
  • Cluster epochs and per-shard leadership terms
  • Lease-backed health-check failover (promoteFailedPrimaries)
  • Explicit shard migration with migrateShard
  • Rolling topology: joinNode · leaveNode · promoteFollower

Search planner

  • Replica-ok reads with primary fallback
  • Pin to one shard with routingKey
  • Prune fan-out with shardIndexes
  • maxReplicaLag filters stale followers when requested
  • Bounded fan-out: maxConcurrentShards, timeoutMs
  • Partial-result merges and cancellation via signal

Security

  • Bearer-token authentication
  • Optional scoped tokens: readToken · writeToken · adminToken
  • TLS deployment guidance for network nodes

Observability

  • Prometheus /metrics for write, search, and durability
  • Latency histograms for queries and durable writes
  • Cluster epoch, log index, and replication progress gauges
  • stats() · health() · durability() APIs
  • Optional OpenTelemetry tracer hook

Operations

  • flush to durably persist outstanding writes
  • optimize to consolidate and trim the index
  • unload / load / isLoaded for cold-storage switching
  • Background maintenance loop on network nodes

Developer API

  • Small surface: add · addMany · delete · deleteMany
  • Search: search · searchMany with stable result shape
  • Upsert semantics on every write
  • Reads: has · get · getMany · getMetadata · setMetadata
  • IDs are string or number
Limited engagements · select partners only

Built to be licensed, not run by someone else.

The capability surface above is the real shipping API. Run a private benchmark on your own data, then license or acquire the engine.

Request access Try the public demo